Evaluation of the methods used by NVE for flood frequency estimation
Donna Wilson
2Outline
Introduction
Flood frequency analysis
Data
Selection of the statistical distribution
Assumption of stationarity
Rainfall runoff modelling
PQRUT
Rainfall inputs
Snowmelt
Final comments
3Introduction
Flood frequency estimation is important for dam design, flood defence schemes and spatial planning.
Property, health and lives are at risk if dams or defence schemes fail to perform to the intended standard or if flood risks are ignored.
Flood estimation is a difficult task, particularly for long return periods.
Two methods available:
Flood frequency analysis (statistical method)
Rainfall-runoff modelling
4 A statistical approach for estimating flood frequency characteristics based on observed data.
A flood frequency curve plots magnitude versus return period. A flood frequency curve is constructed as the product of the index flood and the
growth curve.
Focus on: (1) data(2) selection of the statistical distribution(3) assumption of stationarity
Flood frequency analysis
Index flood = mean/median of the annual maxima flood series.
A growth curve represents how floods increase at longer return periods.
5Data
Where long records of data are available, in excess of the required return period, flood frequency analyses can be relatively straightforward.
..Unfortunately, data are often not available for the site of interest.
Flood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
6NVEs recommended procedures based on available data
Recommended procedures for deriving the index flood and growth curve based onavailable station data
Flood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
Data available Procedure for calculation of the index flood
Procedure for calculation of growth curve for target return periods between Q200 and Q1000
> 50 years Calculated from observed series Calculated from 2- or 3-parameter distribution, based on observed series
30-50 years Calculated from observed series Calculated from 2- parameter distribution, based on observed series
10-30 years Calculated from observed series Calculated by analysis of other long series in the area
< 10 years Calculated by correlation with other series and/ or from flood formulas.
Calculated by analysis of other long series in the area
Ungauged site in an ungauged catchment
Comparison with nearby sites or calculation from formulas
Use of regional flood frequency curves
7212
181
Effect of using short records(Lakshola, Northern Norway)
Flood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
207
246
This short record underestimates the:
200 yr flood by 26m3s-1 (14%)
1000 yr flood by 34m3s-1 (16%)
relative to the longer record
8Regional analysis
Regional growth curves (NVE, 2009).
Weaknesses :
-Variations in catchment characteristics and climate can lead to different flood responses.
- Greater use could be made of observed data.
Strengths:
-Generates a flood estimate in the absence of site data.
- More accurate estimates can be obtained using regional analysis, compared to at-site analysis with limited data.
Flood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
Annual flood(K1 and K2)
SpringSummer and autumn
Flood regions
Spring
Summer & autumn
9Regional analysis: pooling data a better approach?
The regional approach used in some countries (e.g. UK, Germany, Italy, Slovakia) involves the pooling of station data.
5T station-years of data required.
Not necessary to use fixed regions.
Further research would be required to establish the suitability of such an approach for Norwegian sites.
Flood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
10
Selection of the statistical distributionFlood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
The best distributions for Norwegian data are often: Gumbel or Log-Normal (2-parameter distributions) General Extreme Value (3 -parameter distribution)
Selected distribution provides best fit to data.
11
Krinsvatn (Central Norway)
200 yr floodDifference of 69m3s-1
(301 m3s-1 12%)
1000 yr floodDifference of 134m3s-1
(368 m3s-1 18%)
Flood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
Uncertainty increases with increasing return periods
12
3-parameter distributions can be very sensitive to outliers.
Some countries (e.g. UK, Italy) recommend the use of a particular distribution.
NVEs guidelines are flexible, but:
specify minimum periods of record for use of 3-parameter distributions (i.e. 50 years)
recommend that several different distributions are compared.
These are key strengths of the approach recommended by NVE.
A default distribution increases consistency between analysts, but could severely under or over-estimate flood magnitudes.
Flood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
Comments regarding selecting a statistical distribution
13
Assumption of stationarity
Current methods assume that data are stationary.
Environmental changes (e.g. climate change, urbanisation, extensive tree clearing) can lead to major changes in flood frequency.
Two main considerations are that:
(1) past observations may not be stationary
(2) flood frequencies in the future may not be stationary
If a data series has a trend, flood estimates may give a poor representation of current or future flood frequencies.
Flood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
14
Past trends in the spring flood
Flood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
1921 - 2005 1941 - 2005 1961 - 2000
Timing
Magnitude
Wilson, D., Hisdal, H., Lawrence, D. (2010) Has streamflow changed in the Nordic countries? Recent trends and comparisons to hydrological projections. Journal of Hydrology, 394, 334-346.
15
Ensemblemedian
Change (%)
Ensemble90th percentile
Change (%)
Projected changes in 200-year flood (between 1961-1990 and 2071-2100)
Flood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
Lawrence, D (2010) Hydrological projections for changes in flood frequency under a future climate in Norway and their uncertainties. In: Hydrology: From research to water management, NHP Report No. 51, 203-204.
16
Comments regarding stationarity
New approaches are needed for the analysis of non-stationary series.
Past trends: Timeseries from some stations have been analysed for trends, but
this information is not considered in routine flood frequency analysis.
Future changes: Currently no guidance for dealing with projected environmental
change Results from climate change projections are being used to develop
guidance for incorporating the effects of climate changes into flood estimates.
Flood frequency analysis: (1) Data, (2) Selection of the statistical distribution, (3) Assumption of stationarity
17
Rainfall-runoff modelling
A rainfall input (for a particular return period) is converted to a flow output using a model of the catchment response.
A simple, lumped, event based precipitation model (PQRUT) is used.
The method and computer program for this model were developed in the 1980s and are still in use with few modifications.
Focus on:
(1) PQRUT
(2) rainfall inputs
(3) snowmelt
18
PQRUT Model parameters are
calculated from equations based on catchment descriptors or by calibration against observed flows.
Model calibration is rarely achieved using observed data.
Can be run for any time resolution. The PQRUT model
Faster rate
Slower rate
K1 = 0.0135 + 0.00268*HL 0.01665 * ln (ASE)
K2 = 0.009 + 0.21*K1 0.00021*HL
T = -9.0 + 4.4*K1-0.6 + 0.28*QN
Constants for calculation of faster and slower flow rates
Threshold level
Rainfall-runoff modelling: (1) PQRUT, (2) Rainfall inputs, (3) Snowmelt
H = accumulated rainfall and snowmelt
19
PQRUT
Equations used to derive the model parameters were developed for 20 catchments with relatively small catchment areas (
20
Comments regarding PQRUT Simplified version of the HBV model - simple and quick to use.
In Norway there are over 2000 (~ 300 highest class) dams which are subject to review every fifteen years.
High-resolution precipitation and discharge data are being used to improve model output at a series of test sites.
It would be informative to compare the performance of PQRUT against:
the full scale HBV model
alternative event-based rainfall-runoff modelling methods
newer approaches such as continuous simulation modelling.
NVE are involved in two projects comparing the rainfall-runoff model used in Norway with those used in other European countries.
Rainfall-runoff modelling: (1) PQRUT, (2) Rainfall inputs, (3) Snowmelt
21
Areal rainfall
Estimates of extreme rainfall are required (e.g. 200 year, 1000 year, PMP).
Point rainfalls are only representative of a very small area. Average rainfall over a catchment is likely to be much smaller.
Aerial reduction factors (ARFs) are used to account for the effect of space and time variations.
Met.no plans to reassess estimates of extreme precipitation, using grid based data observations. ARF as a function of catchment size and
storm duration (Frland, 1992).
Rainfall-runoff modelling: (1) PQRUT, (2) Rainfall inputs, (3) Snowmelt
22
Rainfall profile
The design storm depth is distributed with a design storm profile.
The distribution can be symmetrical, skewed and/or peaked, but the rainfall profile does not necessarily reflect typical catchment conditions.
Ideally the storm profile will be representative of the typical storm profile (if such a profile exists).
Rainfall-runoff modelling: (1) PQRUT, (2) Rainfall inputs, (3) Snowmelt
23
Snowmelt In many parts of Norway, flood events are generated by a combination of
both extreme precipitation and simultaneous snowmelt.
PQRUT has a simple routine for estimating snowmelt
S = Cs * TL
Where: S = snowmelt in mm/dayCs = degree day factor in mm/C/24 hours (varies depending on presence
of rainfall and dominant land use)TL = air temperature
Snowmelt is added as a fixed amount in the PQRUT model.
This approach is conservative, generally overestimating flood magnitudes.
What snowmelt event should be combined with a 1000 year rainfall event (or less) to generate a 1000 year flood event?
Rainfall-runoff modelling: (1) PQRUT, (2) Rainfall inputs, (3) Snowmelt
24Frequency curves for peak flows resulting from rain-on-snow and rainfall events for a Canadian catchment (Harr, 1981)
Generating mechanisms for a peak flow of 10 l/s per ha:
Flood is 5 times more likely to result from rain-on-snow than rain alone
Rainfall-runoff modelling: (1) PQRUT, (2) Rainfall inputs, (3) Snowmelt
-Peak caused by rain alone has a return period of 15 years
-Peak caused by rain-on-snow has a return period of only 3 years
25
Comments regarding snowmelt
A greater understanding of the combined incidence of rainfall and snowmelt is required.
Rainfall-runoff modelling: (1) PQRUT, (2) Rainfall inputs, (3) Snowmelt
26
Final comments Focussed on 3 issues related to flood frequency analysis, and 3 issues
related to rainfall-runoff modelling. There are other issues which it has not been possible to discuss.
The procedures used by NVE are robust, with the rainfall-runoff method providing conservative flood estimates. However, the procedures need to be subject to continual review and development.
NVE are currently:
Developing procedures for including projected climate change
Improving PQRUT parameter estimation
Met.no plan to reassess estimates of extreme precipitation.
NVE are also working with various European partners to compare and evaluate methods, particularly:
regional approaches
selection of statistical distributions
rainfall-runoff model performance
inclusion of snowmelt
Evaluation of the methods used by NVE for flood frequency estimationOutlineIntroductionFlood frequency analysisDataNVEs recommended procedures based on available dataEffect of using short records(Lakshola, Northern Norway)Regional analysisRegional analysis: pooling data a better approach?Selection of the statistical distributionKrinsvatn (Central Norway)Comments regarding selecting a statistical distributionAssumption of stationarityPast trends in the spring floodLysbildenummer 15Comments regarding stationarityRainfall-runoff modellingPQRUTPQRUTComments regarding PQRUTAreal rainfallRainfall profileSnowmeltLysbildenummer 24Comments regarding snowmeltFinal comments